Bidirectional Stereo Disparity from Professional Stereoscopic Cinema

[DepthAuteur]

Hi there @nagadomi and all other enthusiastic 3D-folks,

I've been working on a fine-tuned S²M² model that produces both negative and positive disparity — trained exclusively on professionally produced native stereoscopic films spanning 70 years of cinema history (1953–2018).

Every existing stereo matching and monocular depth model outputs positive-only disparity. S²M²cinematic is the first public model that learns where objects belong in front of the screen plane (negative parallax) versus behind it, directly from the creative decisions of professional stereographers.

Results after 18 epochs (all achieved best validation ★): Sign-Loss reduced 4.2×, asymmetry ratios 0.95–1.00 on reference frames verified against a passive 3D monitor.

Code and full results: https://github.com/DepthAuteur/s2m2cinematic

No weights released (training corpus is copyrighted film material). The model serves as the ground truth engine for a planned 2D→3D conversion pipeline called Creative Cinematic Stereographing, in short CCS which is already prepared and ready for training :-). GitHub will follow.

Feedback welcome — especially from anyone working on stereoscopic conversion :-)

DepthAuteur

[KolaKater]

it sounds very attractive, could you please post some SBS examples ?

[DepthAuteur]

Thanks for the interest, @KolaKater!

Important context: S²M²cinematic is the ground truth engine for a larger pipeline, not the final conversion model itself. It takes native stereoscopic film pairs (left + right eye) and produces bidirectional disparity maps — those maps then serve as training targets for the actual 2D→3D conversion model (called CCS), which is currently in preparation.

So right now there are no SBS outputs to show — S²M²cinematic's job is to analyze existing stereo films, not to generate new 3D from 2D. The fSBS results will come once CCS training is complete, and I definitely plan to share those.

Stay tuned, CCS GitHub will follow :-)

Best wishes... DepthAuteur

[KolaKater]

thank you for the rapid answer, I have just recognized this: S2M2 is just useless for me. Thanks anyway